Near zero-cost super().meth() calls via adaptive superinstructions

[Fidget-Spinner]

BPO	46564
Nosy	@markshannon, @corona10, @Fidget-Spinner, @kumaraditya303
PRs	bpo-46564: Optimize super().meth() calls via adaptive superinstructions #30992 bpo-46564: do not create frame object for super object #31002

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = 'https://github.com/Fidget-Spinner'
closed_at = None
created_at = <Date 2022-01-28.16:36:58.296>
labels = ['interpreter-core', '3.11', 'performance']
title = 'Near zero-cost super().meth() calls via adaptive superinstructions'
updated_at = <Date 2022-03-29.17:49:41.787>
user = 'https://github.com/Fidget-Spinner'

bugs.python.org fields:

activity = <Date 2022-03-29.17:49:41.787>
actor = 'corona10'
assignee = 'kj'
closed = False
closed_date = None
closer = None
components = ['Interpreter Core']
creation = <Date 2022-01-28.16:36:58.296>
creator = 'kj'
dependencies = []
files = []
hgrepos = []
issue_num = 46564
keywords = ['patch']
message_count = 6.0
messages = ['412005', '412050', '412051', '412052', '412055', '412259']
nosy_count = 4.0
nosy_names = ['Mark.Shannon', 'corona10', 'kj', 'kumaraditya']
pr_nums = ['30992', '31002']
priority = 'normal'
resolution = None
stage = 'patch review'
status = 'open'
superseder = None
type = 'performance'
url = 'https://bugs.python.org/issue46564'
versions = ['Python 3.11']

[Fidget-Spinner]

super().meth() is expensive. I propose to optimize 3 parts of this:

1. Avoid creating a temporary super() proxy object.
2. Avoid creating a bound method.
3. Avoid method lookup in super MRO using the inline cache.

Credit for 1. and 2. doesn't belong to me. Those were inspired by the excellent work done in bpo-43563.

I'll do this by combining the adjacent CALL (super) and LOAD_METHOD instructions into CALL_NO_KW_SUPER__LOAD_METHOD. Using the specializer means:

• We don't touch any compiler code.
• This custom instruction isn't revealed to the user
• I can make use of the 5 cache entries shared by both CALL_ADAPTIVE and LOAD_METHOD_ADAPTIVE.

The final 2-argument super(type, obj).meth() form will have almost no overhead over a corresponding self.meth() call in the current implementation.

Please see faster-cpython/ideas#242 and faster-cpython/ideas#239 for more info.

[kumaraditya303]

I was reading typeobject.c and noticed that creating a super object currently requires creating a frame object which is created lazily and is slow and it would work with the InterpreterFrame as well so I created a PR for this optimization and now it does not requires creating frame objects.

[Fidget-Spinner]

@Kumar, my PR already has your changes.

[kumaraditya303]

Oh, I didn't see your PR and commented as it was not mentioned in this bpo. Would you like to split the PR or continue with yours, either way is fine?

[Fidget-Spinner]

Oh, I didn't see your PR and commented as it was not mentioned in this bpo.

No problem. In the future please check the "Pull Requests" section on the issue. People don't always say "I created a PR at xxx". Often times we just link the PR into the issue itself.

Keep your PR around in case mine gets rejected, then we can take just the good parts. If mine's accepted, we can close yours.

[Fidget-Spinner]

New changeset b9ebde8 by Kumar Aditya in branch 'main':
bpo-46564: do not create frame object for super object (GH-31002)
b9ebde8

[Fidget-Spinner]

Superseded by #87729.